Ai Federated Learning

Federated Optimisation Theory

Doctoral work examines the mathematical foundations of learning distributed across many participants. Theory establishes what guarantees are achievable when data never leaves its origin.

Convergence Analysis Methods

Research establishes conditions under which distributed training reaches a solution. Convergence guarantees determine whether an approach can be trusted at all.

Federated Averaging Algorithms

Doctoral study examines the foundational algorithm combining locally trained models. Averaging remains the baseline against which every method is compared.

Local Computation Strategies

Research examines how much training each participant should perform between rounds. More local work reduces communication but increases divergence between participants.

Server Side Aggregation Methods

Doctoral work examines how contributions from participants are combined centrally. Aggregation design determines robustness, fairness and convergence together.

Adaptive Aggregation Techniques

Research develops combination methods adjusting to observed participant behaviour. Adaptive combination outperforms uniform averaging under realistic conditions.

Proximal Regularisation Methods

Doctoral study constrains how far local models may diverge from the shared model. Constraint improves stability where participant data differs substantially.

Variance Reduction Techniques

Research reduces noise arising from partial participation and local sampling. Lower variance accelerates convergence and improves final model quality.

Momentum Methods In Federation

Doctoral work adapts momentum based optimisation to the distributed setting. Momentum behaves differently when participants train independently between rounds.

Learning Rate Scheduling

Research examines how step sizes should change across training rounds. Scheduling choices strongly influence both stability and final accuracy.

Client Drift Analysis

Doctoral study examines divergence of local models during independent training. Divergence between participants is the central obstacle in federated optimisation.

Objective Inconsistency Research

Research examines mismatch between local and global training objectives. Mismatch can cause training to converge toward an unintended solution.

Partial Participation Analysis

Doctoral work examines training where only some participants contribute each round. Partial participation is universal in practice and complicates all analysis.

Stochastic Participation Modelling

Research models the unpredictable availability of participating devices. Availability patterns correlate with characteristics that bias the resulting model.

Asynchronous Federated Algorithms

Doctoral study examines training without waiting for all participants to respond. Asynchrony suits populations with unpredictable connectivity and availability.

Semi Synchronous Coordination

Research examines coordination schemes between fully synchronous and asynchronous. Intermediate schemes balance efficiency against analytical tractability.

Staleness Handling Methods

Doctoral work addresses contributions computed against outdated shared models. Stale contributions can actively harm rather than assist training progress.

One Shot Federated Learning

Research examines building shared models from a single communication round. Single round methods suit settings where repeated coordination is impossible.

Few Round Federated Methods

Doctoral study minimises the number of coordination rounds required. Round count dominates cost where communication is expensive or restricted.

Federated Distillation Approaches

Research shares predictions rather than model parameters between participants. Sharing predictions greatly reduces communication and permits differing architectures.

Ensemble Based Federation

Doctoral work combines separately trained participant models into an ensemble. Ensembles avoid the averaging problems that heterogeneous data creates.

Model Agnostic Federation

Research develops collaboration independent of the model structure used. Independence permits participants to retain proprietary model designs.

Heterogeneous Model Architectures

Doctoral study enables participants to train structurally different models. Architectural freedom accommodates devices with very different capabilities.

Federated Transfer Learning

Research transfers knowledge across participants with differing feature spaces. Transfer enables collaboration where data is not directly comparable.

Federated Multi Task Learning

Doctoral work treats each participant as solving a related but distinct task. Task relatedness is a more honest description than assuming shared objectives.

Federated Meta Learning

Research learns initialisations that adapt rapidly to each participant. Meta learning connects federated training with effective personalisation.

Federated Reinforcement Learning

Doctoral study examines policy learning distributed across separate environments. Distributed policy learning suits robotics, control and recommendation settings.

Federated Unsupervised Learning

Research examines learning structure from distributed data without labels. Most data held by participants is entirely unlabelled in practice.

Federated Self Supervised Methods

Doctoral work learns representations from unlabelled distributed data. Self supervision exploits the abundant unlabelled data participants already hold.

Federated Semi Supervised Learning

Research combines scarce labelled data with abundant unlabelled data across participants. Labels are typically concentrated among a very small number of participants.

Federated Representation Learning

Doctoral study learns shared representations across distributed participants. Shared representations transfer better than shared task specific models.

Federated Clustering Methods

Research groups distributed data without centralising any of it. Clustering supports both analysis and structuring of participant populations.

Federated Dimensionality Reduction

Doctoral work computes compact representations across distributed datasets. Reduction supports analysis where raw data cannot be assembled centrally.

Federated Matrix Factorisation

Research factorises distributed interaction data without pooling it. Factorisation underpins recommendation across many separate data holders.

Federated Graph Learning

Doctoral study learns over graphs distributed across separate holders. Graph structure frequently spans organisational boundaries that cannot be crossed.

Federated Recommendation Systems

Research builds recommendation without collecting individual interaction histories. Interaction histories are among the most privacy sensitive data collected.

Federated Time Series Modelling

Doctoral work models temporal data distributed across many separate sources. Temporal data is frequently continuous, personal and impossible to centralise.

Federated Natural Language Modelling

Research trains language models on distributed text without collecting it. Personal text is highly sensitive and legally difficult to centralise.

Federated Computer Vision

Doctoral study trains visual models across distributed image collections. Images frequently contain identifiable people and cannot be freely shared.

Federated Multimodal Learning

Research combines several data types held by differing participants. Participants frequently hold complementary rather than identical data types.

Non Identical Data Distribution Handling

Doctoral work addresses participants holding statistically different data. Distribution differences are the norm and degrade naive federated methods.

Label Distribution Skew Research

Research examines participants holding very different class proportions. Extreme skew causes local models to diverge in incompatible directions.

Feature Distribution Skew Research

Doctoral study examines systematic differences in input characteristics. Feature differences arise from devices, locations and collection practices.

Quantity Skew Analysis

Research examines participants holding vastly different data volumes. Volume imbalance causes large holders to dominate the resulting model.

Concept Shift Across Clients

Doctoral work examines cases where identical inputs imply different outputs. Concept differences make a single shared model fundamentally inappropriate.

Distribution Shift Detection

Research detects when participant data changes over the course of training. Undetected shift silently degrades models already deployed in use.

Domain Adaptation In Federation

Doctoral study adapts shared models to differing participant conditions. Adaptation recovers performance where a single model serves poorly.

Personalisation Methods

Research balances a shared global model against individual customisation. Personalisation addresses diversity that any single model serves badly.

Local Fine Tuning Approaches

Doctoral work adapts the shared model locally after federated training. Local adaptation is simple yet frequently very effective in practice.

Parameter Decoupling Methods

Research separates shared parameters from those kept private locally. Decoupling permits personalisation without abandoning collaboration entirely.

Clustered Federated Learning

Doctoral study groups similar participants and trains a separate model per group. Grouping resolves conflicts that a single shared model simply cannot.

Client Similarity Estimation

Research estimates similarity between participants without inspecting their data. Similarity estimation underpins grouping and selective collaboration.

Mixture Of Experts Federation

Doctoral work routes participants toward specialised model components. Specialisation accommodates heterogeneity without fragmenting collaboration.

Hypernetwork Based Personalisation

Research generates participant specific models from a shared generator. Generation permits personalisation while sharing most learned knowledge.

Meta Learning For Personalisation

Doctoral study learns models designed to adapt quickly to each participant. Rapid adaptation matters where participants hold very little data.

Fairness Across Clients

Research examines whether models serve all participants comparably well. Aggregate accuracy routinely conceals very poor service for some participants.

Performance Disparity Analysis

Doctoral work measures how performance varies across participating populations. Disparity measurement is a prerequisite for addressing unfairness at all.

Minority Client Protection

Research protects participants whose data differs from the majority. Minority participants are otherwise systematically disadvantaged by averaging.

Device Capability Heterogeneity

Doctoral study addresses participants with very different computing capacity. Capability differences cause slow devices to hold back the entire process.

Computation Budget Adaptation

Research adapts local work to each participant available resources. Adaptation keeps weak devices contributing rather than excluding them entirely.

Model Partitioning For Weak Devices

Doctoral work divides models so constrained devices can still participate. Partitioning broadens participation to devices that could not otherwise train.

Submodel Extraction Methods

Research assigns smaller portions of a model to less capable participants. Extraction lets a single collaboration span very diverse hardware.

Adaptive Model Complexity

Doctoral study varies model size according to participant capability. Complexity adaptation balances inclusion against achievable model quality.

Client Availability Modelling

Research models when participants are able and willing to contribute. Availability correlates with time zone, device and socioeconomic factors.

Client Departure Handling

Doctoral work addresses participants leaving partway through a training round. Departure mid round is routine and must not corrupt the process.

Straggler Mitigation Strategies

Research addresses participants responding far more slowly than others. Slow participants otherwise determine the pace of the entire collaboration.

Client Selection Strategies

Doctoral study determines which participants contribute in each round. Selection influences convergence speed and representational fairness together.

Sampling Bias In Client Selection

Research examines bias introduced by non random participant selection. Selection bias produces models that serve unselected populations poorly.

Participation Incentive Design

Doctoral work examines why participants would contribute their resources. Participation consumes energy and bandwidth that owners must be willing to spend.

Contribution Measurement Methods

Research quantifies how much each participant improved the shared model. Measurement underpins both fair reward and detection of free riding.

Data Valuation In Federation

Doctoral study values participant data without directly inspecting it. Valuation is required wherever collaboration involves commercial exchange.

Reward Allocation Mechanisms

Research designs distribution of benefit among contributing participants. Perceived unfairness in allocation causes collaborations to collapse.

Federated Learning Marketplaces

Doctoral work examines platforms coordinating collaboration between organisations. Marketplace design determines who participates and on what terms.

Coalition Formation Analysis

Research examines which participants benefit from collaborating together. Not every possible collaboration improves outcomes for its members.

Game Theoretic Analysis Of Federation

Doctoral study models participants as strategic actors pursuing their own interests. Strategic analysis predicts behaviour that cooperative assumptions miss.

Privacy Threat Modelling

Research characterises what an adversary could learn from federated systems. Threat modelling establishes what privacy claims can honestly be made.

Gradient Inversion Attack Research

Doctoral work examines reconstructing training data from shared learning signals. Reconstruction attacks show that keeping data local is insufficient alone.

Membership Inference Analysis

Research examines determining whether specific records participated in training. Membership disclosure is itself a serious privacy harm in many settings.

Property Inference Attacks

Doctoral study examines inferring population characteristics from shared models. Aggregate properties can be commercially and personally sensitive.

Reconstruction Attack Defence

Research develops defences against recovery of data from shared signals. Defence design must preserve enough signal for learning to succeed.

Differential Privacy Mechanisms

Doctoral work applies formal privacy guarantees within distributed training. Formal guarantees replace informal assurances that repeatedly prove inadequate.

Privacy Budget Management

Research examines allocating and tracking cumulative privacy expenditure. Budgets deplete across rounds and must be managed across whole deployments.

Local Differential Privacy Methods

Doctoral study applies privacy protection before anything leaves a device. Local protection requires no trust in the coordinating server at all.

Central Differential Privacy Analysis

Research examines protection applied during central aggregation. Central protection gives better accuracy but requires trusting the aggregator.

Privacy Utility Trade Off Analysis

Doctoral work characterises the cost of privacy protection in model quality. Explicit trade off analysis replaces implicit and undocumented compromise.

Privacy Accounting Methods

Research develops tighter accounting of accumulated privacy loss. Tighter accounting permits stronger models at equivalent protection levels.

Secure Aggregation Protocols

Doctoral study combines contributions without revealing any individually. Secure aggregation is what makes federation genuinely privacy protecting.

Cryptographic Aggregation Efficiency

Research reduces the cost of cryptographically protected combination. Cryptographic overhead currently limits deployment at very large scale.

Homomorphic Encryption Methods

Doctoral work performs computation directly upon encrypted contributions. Encrypted computation removes exposure even during the aggregation stage.

Secure Multiparty Computation

Research enables joint computation where no party reveals its inputs. Multiparty protocols permit cooperation between mutually distrusting organisations.

Trusted Execution Environments

Doctoral study examines hardware isolated computation within federated systems. Hardware isolation offers protection with far lower computational overhead.

Hardware Attestation Methods

Research verifies that participants are running the expected software. Attestation prevents participants misrepresenting what they actually executed.

Zero Knowledge Verification

Doctoral work proves properties of contributions without revealing them. Zero knowledge proofs reconcile verification with strong confidentiality.

Verifiable Computation Methods

Research verifies that participants performed the work they claim. Verification prevents both free riding and deliberate corruption of training.

Model Poisoning Defence

Doctoral study defends against deliberately corrupted participant contributions. Open participation exposes training to motivated interference.

Data Poisoning Detection

Research detects corrupted training data held by individual participants. Detection is difficult because the underlying data is never directly observable.

Backdoor Attack Research

Doctoral work examines hidden triggers implanted within shared models. Implanted triggers remain dormant through ordinary testing and evaluation.

Backdoor Detection Methods

Research develops identification of hidden malicious behaviour in models. Detection must operate without access to the data that created it.

Byzantine Robust Aggregation

Doctoral study designs combination resistant to arbitrarily corrupted participants. Robust combination tolerates participants behaving in entirely unpredictable ways.

Robust Statistics In Aggregation

Research applies statistically robust estimators to combining contributions. Robust estimators resist corruption but may weaken convergence guarantees.

Anomaly Detection In Contributions

Doctoral work identifies participant contributions departing from expected patterns. Anomalies may indicate corruption, malfunction or genuine unusual data.

Sybil Attack Mitigation

Research addresses adversaries creating many false participant identities. False identities can outvote genuine participants within robust schemes.

Free Rider Detection

Doctoral study identifies participants benefiting without genuinely contributing. Free riding undermines the incentives sustaining a collaboration.

Collusion Resistance Analysis

Research examines groups of participants cooperating to subvert a federated system. Collusion defeats defences designed only against isolated adversaries.

Adversarial Robustness In Federation

Doctoral work examines model behaviour under deliberately manipulated inputs. Robustness must be achieved without centralised access to training data.

Model Extraction Defence

Research protects shared models from reconstruction by participants. Every participant receives the model and could attempt to appropriate it.

Intellectual Property Protection

Doctoral study examines ownership of models built from many contributions. Ownership uncertainty inhibits collaboration between commercial organisations.

Watermarking Federated Models

Research embeds ownership markers within collaboratively trained models. Watermarks support claims where models are copied without permission.

Model Provenance Tracking

Doctoral work records which contributions shaped a given deployed model. Provenance is required for both accountability and regulatory review.

Audit Trail Design

Research designs records of what a federated system did and when it did so. Audit records support investigation without exposing any participant data.

Accountability Frameworks

Doctoral study examines responsibility when collaboratively built models cause harm. Distributed development distributes and obscures responsibility considerably.

Trust Establishment Mechanisms

Research examines how mutually unfamiliar parties come to collaborate. Trust establishment is the practical precondition for any federation.

Reputation Systems

Doctoral work develops reputation tracking across repeated collaborations. Reputation supports selection where direct verification is impractical.

Blockchain Based Coordination

Research examines distributed ledgers coordinating federated collaboration. Ledgers offer auditability without any single trusted coordinator.

Decentralised Trust Architectures

Doctoral study removes reliance on any single trusted coordinating party. Decentralisation eliminates a single point of both failure and control.

Regulatory Compliance Analysis

Research examines demonstrating compliance across distributed participants. Evidence collection is difficult when data never reaches a central system.

Data Protection Law Analysis

Doctoral work examines how data protection law applies to federated systems. Legal treatment of model parameters remains substantially unsettled.

Cross Border Data Governance

Research examines federation spanning differing legal jurisdictions. Federation is frequently proposed precisely to satisfy residency requirements.

Consent Management In Federation

Doctoral study examines recording and honouring participant permissions over time. Consent achieves nothing unless the system technically enforces it.

Right To Erasure Research

Research examines honouring requests to remove data influence from models. Removal is legally required yet technically very difficult to achieve.

Federated Unlearning Methods

Doctoral work removes the influence of specific data from a trained model. Retraining entirely from scratch is usually prohibitively expensive.

Verification Of Data Removal

Research verifies that requested data removal genuinely took effect. Verification is essential because removal claims cannot easily be checked otherwise.

Data Minimisation Principles

Doctoral study examines transmitting only what a purpose genuinely requires. Minimisation reduces privacy exposure and communication cost together.

Purpose Limitation Analysis

Research examines restricting model use to the purposes originally agreed. Models trained for one purpose are very readily repurposed for others.

Privacy Impact Assessment Methods

Doctoral work develops structured assessment of federated system privacy risk. Assessment frameworks designed for centralised systems fit poorly here.

Communication Efficient Methods

Research reduces data exchanged between participants and coordinators. Communication rather than computation dominates federated training cost.

Gradient Compression Techniques

Doctoral study compresses learning signals transmitted during federated training. Compression makes federation viable over severely constrained network links.

Quantisation Of Contributions

Research reduces numerical precision of transmitted training signals. Lower precision reduces transmission volume with modest accuracy cost.

Sparsification Methods

Doctoral work transmits only the most significant portions of learning signals. Most transmitted values contribute negligibly to eventual model quality.

Sketching And Sampling Approaches

Research summarises contributions using compact randomised representations. Sketches preserve aggregate information at a fraction of the size.

Error Feedback Mechanisms

Doctoral study accumulates and later transmits information lost to compression. Feedback recovers accuracy that aggressive compression would otherwise cost.

Bandwidth Aware Scheduling

Research schedules communication according to available network capacity. Scheduling avoids congestion that would otherwise stall the collaboration.

Network Topology Effects

Doctoral work examines how connection structure influences training behaviour. Topology determines both convergence speed and communication burden.

Hierarchical Federated Architectures

Research aggregates through intermediate tiers before central combination. Hierarchy reduces the communication reaching central infrastructure.

Edge Server Coordination

Doctoral study examines coordination performed at network edge infrastructure. Edge coordination reduces latency and central bandwidth requirements.

Decentralised Peer Learning

Research examines collaboration through direct exchange between participants. Decentralisation removes reliance on any central coordinating server.

Gossip Protocol Methods

Doctoral work examines randomised information spreading between peers. Gossip protocols scale gracefully without any central coordination point.

Topology Optimisation Research

Research designs which participants should communicate with which others. Connection design trades communication cost against information spreading speed.

Wireless Federated Learning

Doctoral study examines federation conducted over unreliable wireless connections. Wireless links lose data routinely and cannot be assumed reliable.

Over The Air Computation

Research exploits wireless signal superposition to aggregate during transmission. Aggregating in the channel itself avoids transmitting each contribution separately.

Mobile Network Integration

Doctoral work integrates federated systems within mobile network infrastructure. Network operators are well placed to coordinate at very large scale.

Energy Efficient Federation

Research reduces energy consumed by participation in distributed training. Energy cost falls on participants and determines willingness to contribute.

Battery Aware Client Scheduling

Doctoral study schedules participation according to available stored energy. Scheduling avoids depleting batteries that owners need for other purposes.

On Device Training Efficiency

Research develops efficient local training on severely constrained hardware. Training is far more computationally demanding than inference on the same device.

Memory Constrained Training

Doctoral work enables model training within very limited memory budgets. Available memory is frequently the binding constraint on participating devices.

Hardware Acceleration For Clients

Research examines specialised hardware supporting local training on devices. Acceleration determines what model complexity participants can realistically support.

Federated Learning Frameworks

Doctoral study examines software platforms implementing federated systems. Framework capability determines what research can practically be attempted.

System Benchmarking Methods

Research develops fair comparison between federated system implementations. Reported comparisons frequently omit communication and energy costs.

Simulation Environment Design

Doctoral work develops environments simulating large federated deployments. Simulation permits evaluation at scales real deployment cannot reach.

Reproducibility In Federated Research

Research establishes practices allowing federated results to be repeated. Participant heterogeneity makes reproduction unusually difficult in this field.

Scalability To Large Populations

Doctoral study examines behaviour as participant numbers grow very large. Approaches viable for hundreds frequently fail across millions of participants.

Cross Device Deployment Research

Research examines federation across very many small consumer devices. This setting combines extreme scale with unreliable and brief participation.

Cross Silo Deployment Research

Doctoral work examines federation between a few large organisations. Organisational federation raises legal and commercial rather than scale challenges.

Production System Engineering

Research examines engineering federated systems for sustained real operation. Production requirements differ substantially from research prototypes.

Model Versioning And Rollout

Doctoral study examines distributing model revisions across participant populations. Staged distribution limits how far a faulty revision can spread.

Deployment Monitoring Systems

Research monitors deployed federated models across participant populations. Monitoring must operate without collecting the data it would need centrally.

Model Degradation Detection

Doctoral work detects declining performance among deployed models. Degradation is difficult to detect where evaluation data is also distributed.

Continual Federated Learning

Research examines federation continuing indefinitely as new data arrives. Continuous operation differs greatly from fixed and bounded training campaigns.

Catastrophic Forgetting In Federation

Doctoral study examines loss of earlier capability during continued training. Forgetting is amplified when participants hold very different data.

Lifelong Learning Across Clients

Research examines accumulating knowledge across a changing participant population. Populations turn over while the shared model must persist.

Data Lifecycle Management

Doctoral work examines how participant data is retained and eventually discarded. Lifecycle rules interact directly with legal retention obligations.

Labelling Strategies In Federation

Research examines obtaining reliable labels from distributed participants. Central label review is impossible when the data itself cannot be centralised.

Label Scarcity Handling

Doctoral study addresses settings where very few labels are available. Labels are typically far scarcer in federated than centralised settings.

Active Learning In Federation

Research selects which distributed examples would most benefit from labelling. Selection must operate without inspecting the candidate data centrally.

Human In The Loop Federation

Doctoral work incorporates human judgement within distributed training. Human involvement must respect the privacy federation was chosen to protect.

Healthcare Federated Learning

Research examines collaboration between healthcare organisations without sharing records. Healthcare is the most frequently cited motivation for this approach.

Medical Imaging Federation

Doctoral study trains imaging models across separate hospital collections. Imaging models require diversity no single institution can provide.

Electronic Health Record Federation

Research examines learning across separate clinical record systems. Record structures differ substantially between institutions and countries.

Genomic Data Federation

Doctoral work examines collaboration across separate genomic data holdings. Genomic data is permanently identifying and rarely permitted to move.

Clinical Trial Data Collaboration

Research examines federation across separately conducted clinical studies. Collaboration increases statistical power without pooling participant records.

Wearable Health Data Federation

Doctoral study examines learning across continuously monitored personal devices. Continuous personal data is both abundant and extremely sensitive.

Financial Services Applications

Research examines collaboration between financial institutions without sharing records. Competitive and regulatory barriers both prevent direct data sharing.

Fraud Detection Federation

Doctoral work examines shared fraud detection across separate institutions. Fraud patterns cross institutions that cannot exchange transaction records.

Credit Risk Model Federation

Research examines collaborative credit assessment across separate lenders. Collaboration raises significant fairness, transparency and accountability obligations.

Insurance Analytics Federation

Doctoral study examines shared modelling across insurance providers. Shared risk modelling raises concerns about coordination and exclusion.

Telecommunications Applications

Research examines federation across telecommunications network operators. Operators hold detailed behavioural data they cannot lawfully pool.

Network Optimisation Federation

Doctoral work applies federation to the management of communication networks. Networks generate distributed data suited naturally to this training approach.

Internet Of Things Applications

Research examines federation across very large populations of connected devices. Device populations are enormous, constrained and intermittently connected.

Smart Home Federated Systems

Doctoral study examines learning across devices within private homes. Household data reveals occupancy, behaviour and personal routine directly.

Industrial Federated Learning

Research examines collaboration between manufacturing sites and organisations. Process data is commercially sensitive and rarely shared openly.

Predictive Maintenance Federation

Doctoral work shares equipment failure knowledge across separate operators. Failures are rare, which makes pooled operational experience unusually valuable.

Autonomous Vehicle Federation

Research examines learning across vehicle fleets without centralising recordings. Vehicle recordings are enormous in volume and capture public spaces.

Traffic And Mobility Federation

Doctoral study examines mobility modelling across separate data holders. Mobility traces are exceptionally identifying even when anonymised.

Smart Grid Federated Analytics

Research examines learning across electrical network measurement points. Consumption data reveals detailed household activity and occupancy patterns.

Agricultural Federated Systems

Doctoral work examines collaboration across farms and agricultural operations. Farm data is commercially sensitive and connectivity is frequently poor.

Retail And Commerce Applications

Research examines federation across retailers and commercial platforms. Purchase data is both competitively valuable and personally revealing.

Keyboard And Input Prediction

Doctoral study examines text prediction trained across personal devices. This application was among the earliest large scale federated deployments.

Speech Recognition Federation

Research trains speech models without collecting personal voice recordings. Voice recordings are inherently identifying and highly sensitive.

Federated Learning In Education

Doctoral work examines federation across separate educational institutions. Learner data is sensitive and very frequently concerns children directly.

Public Sector Applications

Research examines federation between government departments and agencies. Public sector data sharing carries particularly high accountability requirements.

Humanitarian And Development Applications

Doctoral study examines federation in humanitarian and low resource settings. Protecting vulnerable populations makes data protection especially critical.

Scientific Collaboration Federation

Research examines federation between research institutions and consortia. Collaboration enables analysis no single institution could support alone.

Evaluation Methodology

Doctoral work develops sound evaluation of federated learning systems. Evaluation must account for communication, privacy and fairness simultaneously.

Benchmark Dataset Design

Research constructs datasets reflecting genuine federated conditions. Existing benchmarks frequently misrepresent real participant heterogeneity.

Realistic Data Partitioning

Doctoral study examines how benchmark data is divided among simulated participants. Partitioning choices largely determine reported method performance.

Metric Design For Federation

Research develops measures capturing what federated systems should achieve. Average accuracy alone ignores fairness, cost and privacy entirely.

Statistical Significance In Evaluation

Doctoral work examines sound statistical comparison of federated methods. Participant sampling introduces variance frequently ignored in reported results.

Cost And Sustainability Analysis

Research evaluates whether federation delivers value proportionate to its cost. Federation is frequently more costly than centralised training overall.

Carbon Footprint Of Federation

Doctoral study quantifies environmental burden across distributed participants. Distributed training may consume more total energy than centralised training.

Implementation And Adoption Research

Research examines why federated projects succeed or fail in organisations. Barriers are frequently organisational and legal rather than technical.